Maximum Mean Discrepancy
Also known as MMD
Overview
Maximum Mean Discrepancy (MMD) measures how different two distributions are by comparing their representations in a reproducing kernel Hilbert space. It is used in two-sample and drift tests, domain adaptation, generative modeling, and sampling, where detecting or reducing distribution mismatch matters. Common approaches include kernel mean embeddings, U-statistic estimators, random Fourier feature approximations, and gradient-based optimization; adaptive bandwidth and kernel-selection methods aim to improve sensitivity and power. Theory provides strong estimation and convergence guarantees, but choosing a kernel that detects the relevant differences while keeping tests valid and computation scalable remains a central challenge.
Latest papers
38 papers5 papers in the last four weeks, up 67% on the four weeks before. 0.0% of all new papers.
Don't Match the Noise: Distribution Matching under Unknown Measurement Error with an Audit Sample
Sampling Interventional Distributions with Orthogonal MMD Flow
Robust-by-Design Distributional Learning from Contaminated Samples
MiniMMD: Minimization of MMD for Sampling, Inference and Integration
A Jump-Modified Wasserstein Gradient Flow for Maximum Mean Discrepancy with Energy Kernels
On the Optimisation Landscape of Overparameterised Weighted MMD
Beyond Marginal Monitoring: Distributed Joint-Distribution Testing for Data Concept Drift in Large Scale E-Commerce Operations
Representation-Space MMD for Diffusion Language Models
M3-Score: Fidelity, Memorization and Coverage as Separate Axes for Evaluating Generative Radiology Image Models
How fine a change can moments see? A scale law for detecting distribution shift, with a kernel calibration rule
Minimax Lower Bounds of Kernel Discrepancy Estimation: MMD, HSIC, KSD
Distributional Split Criteria for Random Forests: Extensions, Shrinkage, and the Robustness of Mean Splitting
drforest library, whose Rust-backed split search makes broad criterion sweeps inexpensive.